Autonomous Improvements at Scale: Learning Long-Horizon Tasks with Limited Supervision

Abstract: The rapid advancements in deep learning and large models across diverse fields have opened up exciting new opportunities for enhancing robotic perception and control with data-driven methods. However, the true potential of these methods is often hindered by the limitation of the scalability of human supervision. In contrast to computer vision and natural language processing, the robotics realm faces unique challenges in data gathering and utilization, making it difficult to directly replicate previous successes in other domains. In this talk, I will discuss our recent research on enhancing robot's capabilities to tackle long-horizon manipulation tasks through autonomous improvements. I will start by discussing the utilization of procedural content generation to learn robust skills that can handle the variety and uncertainty of the real world. Then I will present a class of methods that train robots to effectively reuse skills learned from prior data for novel long-horizon tasks by learning to generate feasible subgoals. Finally, I will demonstrate how to enable robots to solve various tasks specified by natural language commands without the need of extensive amounts of labels. 

Bio: Kuan Fang is an incoming Assistant Professor of Computer Science at Cornell University (starting Fall 2024). He is currently a Postdoctoral Researcher at the Berkeley Artificial Intelligence Research (BAIR). He received his Ph.D. degree in Electrical Engineering from Stanford University, advised by Fei-Fei Li and Silvio Savarese. His research interests lie at the intersection of robotics, computer vision, and machine learning, with a focus on developing intelligent robots that can solve diverse and complex tasks in unstructured environments using data-driven methods. He is a recipient of the Stanford Graduate Fellowship and the Computing Innovation Fellowship.